Tisane
AIModerate content and run deep text analysis with the Tisane API, detecting abusive content with severity ratings and extracting sentiment. Agents screen user content and escalate borderline cases to a human moderator.
What This Integration Enables
Tisane sits in this category by accident of technology, not of purpose. It shares a language model with the translators, but its job is the opposite of theirs: it is not producing text for people to read, it is producing a judgment about text people have already written. Analyze Text returns abuse entries typed as bigotry, personal attack, profanity, criminal activity, adult only, spam, and more, each carrying a severity of low, medium, high, or extreme, alongside sentiment expressions, named entities, topics, and optional word-level detail. That severity ladder is the whole design. A moderation system with one bit of output forces every case into the same treatment. A moderation system with four rungs lets you automate the ends and staff the middle.
The rest of the surface supports that job rather than competing with it. Extract Plain Text strips HTML, JavaScript, CSS, and JSON so the analysis reads content rather than markup. Detect Language handles mixed-language text by returning a breakdown by character offset, which matters because a lot of abuse arrives code-switched on purpose. Translate Text runs on top of the same semantic analysis, so it handles slang, jargon, and deliberately obfuscated wording rather than throwing up its hands. Compare Named Entities normalizes name order and honorifics for KYC and watchlist matching, and Compare Semantic Similarity scores meaning across languages for deduplication and plagiarism screening.
Without FlowRunner
With FlowRunner
Use Case Scenarios
Community moderation with a graded response
Every message posted to a community space in Slack passes through the agent. It strips markup, calls Analyze Text with snippets and explanations enabled, and acts on severity. Extreme hits are removed and the incident logged. Low findings, the ordinary profanity of ordinary people, are recorded for reporting and nothing else happens. Medium and high findings go to the moderator channel with the exact offending fragment, the abuse type, and Tisane's rationale, so the moderator reads three lines instead of a thread. The agent is not deciding who gets banned. It is making sure the decisions a person does make arrive with the evidence attached.
Multilingual ticket intake
A support ticket arrives in a language nobody on shift reads. The agent calls Detect Language, which returns a breakdown by character offset with confidence per span, so a ticket that switches languages mid-sentence is handled correctly rather than assigned to whichever language happens to appear first. It then translates the body and writes both versions into Zendesk. In the same pass it runs Analyze Text over the original, because a customer who is genuinely furious and a customer who is making a threat are different tickets and should not share a queue. Sentiment sets the priority. Abuse severity sets the escalation path.
Community health reporting
Moderation without measurement drifts. On a schedule, the agent runs Analyze Text across the period's content and writes the topics, sentiment expressions, and abuse categories with their severities into Google Sheets. The result is a record of what the community actually argued about and how hostile it got, which is the input a trust and safety owner needs to change a policy rather than a hunch about whether things feel worse than last quarter.
Human-in-Loop Highlight
Every other connector in this category can be wrong about a sentence. This one can be wrong about a person. A moderation verdict is an action taken against a human being who wrote something, and a false positive is not a typo. It is a real user silenced by a machine, usually one who was quoting the abuse rather than committing it, or writing in a dialect the model reads as hostile, or using a reclaimed word inside the community that reclaimed it. Tisane returns severity as low, medium, high, or extreme precisely so that this does not have to be a single automated yes or no.
FlowRunner agents act at the ends and stop in the middle. Extreme is removed on sight and logged. Low is counted, not actioned. Medium and high stop the flow and become a decision. The agent posts to the moderator queue with the message alongside four things a person needs in order to judge it: the offending fragment surfaced by Include Snippets, the abuse type, the severity, and the explanation Tisane gave. Then it asks: "Personal attack, severity high, in #general from a two-year account with no prior flags. Remove the post, remove and warn, or leave it and mark this a false positive?" That last option is not a courtesy. It is how the policy gets better, because a false positive that nobody records is a false positive that happens again. This is the digital andon cord in its most literal form: the agent pulls the line to a stop the moment the cost of being wrong lands on a person rather than on a process.
Agent Capabilities
8 actionsText Analysis
4- Analyze Text Runs Tisane's full natural-language analysis over a text fragment in a single call. Detects problematic content including hate speech, personal attacks, profanity, sexual advances, criminal activity, adult-only material, self-harm, and spam, each with a severity rating, plus sentiment expressions, named entities, topics, and optional word-level and phrase-structure detail. This is the operation a moderation flow is built around.
- Compare Semantic Similarity Scores how close two text fragments are in meaning, returning a number between 0 and 1. The comparison is semantic rather than lexical, so paraphrases score highly even when they share few words, and the two fragments may be written in different languages. Used for deduplication, plagiarism screening, and matching user input against known phrasings.
- Compare Named Entities Compares two compound personal names and reports whether they refer to the same entity, listing exactly which components differ: given name, surname, title, social role, suffix, or spelling variation. Name order and honorifics are normalized. Used for KYC, deduplication, and watchlist matching.
- Extract Plain Text Strips HTML, JavaScript, CSS, JSON, and other markup from a string and returns decoded plain text, ready to be passed to Analyze Text. Binary content such as PDF or image data is not supported.
Language
2- Detect Language Identifies which languages appear in a text fragment and returns a breakdown by character offset with a confidence score for each span. Supports mixed-language text through a segmentation pattern that splits the input so each segment is identified separately, and the candidate set can be narrowed with a list of expected languages.
- Translate Text Translates a text fragment between any two Tisane-supported languages. Because translation runs on top of Tisane's semantic analysis it handles slang, jargon, and deliberately obfuscated wording, and preserves gender where the target language requires it. Setting the source and target language to the same value paraphrases the text instead of translating it.
Reference Data
2- List Supported Languages Retrieves every language Tisane can process, with its IETF code, native and English names, native encoding, recommended display font, and whether the script is Latin or right-to-left. Used to build language pickers or validate a language code before running an analysis.
- List Inflected Forms Retrieves every inflected form of a lexeme from Tisane's language model, together with the grammatical features of each form. Lexeme and family identifiers come from the word-level output of Analyze Text. Used to build morphological expansions of search terms or blocklist entries.
Frequently Asked Questions
What can FlowRunner do with Tisane?
FlowRunner agents can run Analyze Text, Translate Text, and Detect Language in Tisane, plus 5 more actions.
Does connecting Tisane to FlowRunner require OAuth?
No. Tisane connects to FlowRunner with an API key, no OAuth flow required.
Can Tisane trigger a FlowRunner workflow automatically?
Tisane doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.
Start building with Tisane
$100 in credits. No card required. Connect in minutes.